The Big Four have now all been caught passing off generative content as ‘thought leadership’

The consulting industry’s four biggest players have long used whitepapers and surveys as a means to drum up business – indicating expertise on key specialisms and topical issues their clients are responding to. But in their haste to present as ‘artificial intelligence-led’, KPMG, Deloitte, EY, and now PwC, have potentially alienated customers on this front, by releasing so-called ‘thought leadership’ which has been alleged to be majority-generated by AI.

Thought leadership is one of the key marketing tools of the modern consulting industry. Along with demonstrating niche knowledge in important sectors, offering one or two practical steps potential clients can use to improve their business, and giving insight into consumer or business sentiment via surveys, consultants have used this as a way to encourage businesses to reach out to them for decades.

Amid the excitement of AI investment in recent years, firms have been exploring ways in which the technology might help them to trim costs, and improve their bottom-lines. And as these reports are lead-generating, rather than directly income-generating, there has been a push to make them cost as little as possible, by steadily automating their production.

This rush to cut costs, and demonstrate a sense of being ‘AI-forward’ to investors has come at a cost, when it comes to the credibility of consulting insights, however. In a series of bruising scandals for the industry’s Big Four – KPMG, Deloitte, EY, and PwC – researchers have found that the quartet’s ‘thought leadership’ has become prone to embarrassing ‘hallucinations’ (a euphemism for ‘inaccuracies’).

Damaging trust

Hallucinations aren’t a rare quirk of any AI system, but have rather proven to be a very common occurrence, even after years of development of AI tools. Previous BBC research into leading AI tools found their news summaries contained “significant issues” in the majority of cases, for example. As a result, putting out reems of quick and easy AI prompted content is not a good idea, even in terms of conventional online marketing, where quantity traditionally trumps quality.

Deloitte was the first to find this out the hard way last year, when it was forced to refund part of an Australian government contract – after admitting it used generative AI to produce a report riddled with fake citations, phantom footnotes, and even a made-up quote from a Federal Court judgment. The consulting giant confirmed it would repay AU$440,000 ($291,245) to Australia’s Department of Employment and Workplace Relations, after the department re-uploaded a corrected version of the report, stripping out over a dozen references, and rewriting text – though officials insisted the "substance" of the report remained intact.

KPMG similarly came a cropper when research outfit GPTZero claimed its forensic review of the Big Four firm’s October 2025 report found 40 of the 45 sources were hallucinated by generative tools. ‘Total Experience: Redefining Excellence in the Age of Agentic AI’ also saw half of its factual claims found to be either false, or misattributed. For example, ‘2019 East Japan Press Release’ as evidence of agentic AI usage, despite agentic AI only coming into public discourse in 2024. Meanwhile, the LLM also persistently mistook subjects for authors, meaning articles about certain sources were leant the credibility of being supposedly written by them.

In May 2026, GPTZero also busted EY, alleging that a report on cyber-security was 72% AI-generated. The investigation found that the 2025 study from EY Canada, titled ‘Points of Attack: Uncovering Cyber Threats and Fraud in Loyalty Systems’ was riddled with hallucinations. Some 16 of 27 citations, including some purporting to be from Forbes, McKinsey & Company (itself now a prolific user of generative AI), Gartner, TechCrunch and WIRED “were either broken URLs, or never existed in the first place.”

Now, PwC has capped off the foursome, following another probe by GPTZero – this time into four studies published between 2024 and 2026 for clients in the Middle East. While GPTZero’s own software cannot state for certain, the detection firm said there was an 84% chance that PwC’s report ‘Transforming Governance’ was entirely generative text. Most damaging for PwC, the report promoted a PwC data-driven framework called ‘Citizen Pulse’, claiming it was being used by the governments of Denmark, Saudi Arabia, the United States, and Australia. However, GPTZero found no public evidence that this product was used, nor that these government deals ever took place.

Om Ogale of GPTZero later explained, “Additionally, the report seems unsure if ‘Citizen Pulse’ is an existing PwC product or a vague exercise in thought leadership, cycling between describing the potential benefits of the Citizen Pulse framework and concrete claims about current use. This inconsistency is a common symptom of text generated by a poorly-prompted LLM.”

Machines misquoting machines

PwC did not issue a denial of using generated content in the study in question. Instead, it commented: “PwC Middle East takes the accuracy of our published research seriously and is updating a limited number of supporting citations in the reports mentioned, in a statement eerily similar to that issued by KPMG, in the wake of its own GPTZero study: “KPMG International takes the accuracy and integrity of its published content seriously. The report has been removed and we are reviewing the circumstances surrounding its publication.”

While the two statements are likely too short to confirm whether they too were generated, that is beside the point. The findings erode trust in the ‘thought leadership’ which consultants have traditionally used to drum up business – but also the confidence in their eventual customers that at some point during any engagement, their work is going to be checked and approved by a senior expert, with something on the line. Implying that this ‘quality control’ seems not to apply to this kind of engagement opens up questions of what other services consulting partners may be asleep at the wheel for, or indeed, what clients are actually paying for.

Possibly presenting an even bigger problem, is that ‘hallucinations’ don’t end with the whitepaper they appear in now. For years, companies such as OpenAI and Anthropic have programmed their AI models to seek out ‘reputable’ sources to inform their generative content – and among those reputable sources are the PDF files published by large corporates such as the Big Four, and authoritative journalistic publications (to that end, AI checkers also said KPMG’s study had already appeared in industry publications and one checked newspaper). So when mistakes, hallucinations, inaccuracies which appear in AI-generated ‘thought leadership’ go live, they are primed to be recycled by other LLMs as fact – rendering their ‘insight’ even less credible.

Summing up what this means, Matt Parry – CEO of technology consultancy The Future Collective – warned, “Here is what should worry strategy teams more. These reports do not just sit in a PDF. They are scraped, retrieved and cited by the AI systems that executives now ask for answers, then folded into whatever gets trained next. A product that may never have existed can be repeated back as fact, wearing a trusted domain as credentials. Once a claim has been echoed twice, it looks corroborated. There is no retraction mechanism for any of it. We are planning strategy on that substrate. The frightening part is not that AI wrote the reports. It is that the reports are now writing the AI.”